R4

1
R4

• Andy Aiken

2
Overview

• The Refiner series
• Cognitive psychology of diagnosis
• Diagnosis Web
• Proposed work R4

3
The Refiner Series

• Case-based system
• Cases are assigned to categories
• System generates descriptions of these categories
• Resolves inconsistencies in datasets
• An inconsistency is when a case does not only
  match the category it has been assigned to by an
  expert
• User-directed

4
The Refiner Series

• History
• Refiner initial system (1988)
• Refiner batch system (1995)
• Much more efficient
• Refiner current version
• Java, faster, more strategies and data types, GUI
• Evolution not revolution

5
The Refiner Series
6
The Refiner Series

• Refiner was shown to experts
• Anaesthetics, child psychology, HDU, medical CAL,
  engineering
• Reaction was generally positive
• System to be used in several proposed
  interdisciplinary projects
• BUT

7
The Refiner Series

• Limitations
• Time Series
• Many medical datasets, for example, are
  time-series data
• Noise
• Can deal with noise, but in a limited and
  sometimes confusing way
• This is our focus

8
Refiner and Noise

• What changes are needed?
• Changes the focus of the system
• Refiner the creation of category descriptions
• R4 classification
• This requires us to investigate the science of
  diagnosis

9
Diagnosis

• How do clinicians perform diagnosis?
• How is it taught?
• What is the cognitive psychology?

10
Diagnosis

• Differential Diagnosis
• What is a differential diagnosis?
• Requires the diagnostician to memorise a vast
  amount of information
• Simplistic view of the process

11
Diagnosis
12
Diagnosis

• Hypothetico-Deductive Reasoning
• Uses both forward and backward chaining
• Iterative
• Requires significant cognitive processing

13
Diagnosis

• Given initial facts, work forward to likely
  hypotheses
• These are the differential diagnoses

Facts
Hypotheses
14
Diagnosis

• Identify other facts which should be present for
  these hypotheses to hold, and investigate them
• Can attempt to distinguish between a set of
  competing hypotheses, or get additional
  confirmation for a hypothesis

Facts
Hypotheses
15
Diagnosis

• Repeat until satisfied
• One hypothesis is notably more likely than others
• One likely hypothesis is especially significant
• The courses of action are similar

Facts
Hypotheses
16
Diagnosis

• The effect of expertise
• Novices perform HDR, but the process is slow,
  potentially incomplete and inaccurate
• Intermediates perform HDR, but this is still a
  slow process
• Experts still perform HDR but this is
  supplemented by compiled knowledge (shortcuts
  faster)

17
Diagnosis

• Illness Scripts (Feltovich, 1984)
• Compiled knowledge used in diagnosis
• Comprises three knowledge sources
• Faults
• Consequences
• Enabling conditions
• Says nothing about how they are used

18
Diagnosis

• Faults
• The underlying malfunctions which give rise to
  the symptoms
• Consequences
• Detectable signs and symptoms
• Enabling conditions
• Background knowledge
• Predisposing factors (patients case history)
• Boundary conditions (age and gender)
• Hereditary factors

19
Diagnosis

• Illness scripts dont explain the process of
  diagnosis
• Diagnosis Web
• A more inclusive explanation
• More knowledge sources
• Links between them indicate processes

20
Diagnosis Web
21
Diagnosis Web

• How do we use the diagnosis web to improve the
  Refiner series?

22
R4

• Will be aimed at novice diagnosticians
• Medical students, but not limited to medical
  domains
• An aid to diagnosis
• Supports the students reasoning without taking
  over

23
R4

• Three stories
• An aid to diagnosis
• Scaffolding users diagnostic skill
• Generation of test cases

24
R4

• User presents a case to the system
• The system displays two lists
• Likely (differential) diagnoses
• Potential next steps (descriptors to find values
  for) to hone in on the correct diagnosis
• Providing more data for the case updates these
  lists automatically

25
DXplain
26
DXplain

• Used by clinicians to perform diagnosis
• User enters findings, system updates lists of
  likely diagnoses and other findings to be
  investigated
• Sound familiar?

27
R4 and DXplain (et al)

• Superficially very similar
• R4 is case-based others are database-driven
• Easier to create maintain a casebase
• R4 uses descriptor values others are term-based
• Lends itself to automation
• No other system attempts to mirror the cognitive
  process of diagnosis

28
Conclusion

• Refiner was well received but did not handle
  noisy datasets well
• Adding the facility to handle noisy datasets
  changes the focus to classification
• Combining HDR and illness scripts leads to the
  Diagnosis Web
• The new Refiner system will incorporate aspects
  of HDR and illness scripts to mirror how
  diagnosis is performed

29
Questions?